首页|AutoIHCNet: CNN architecture and decision fusion for automated HER2 scoring
AutoIHCNet: CNN architecture and decision fusion for automated HER2 scoring
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NSTL
Elsevier
In this work, the automated scoring of prognostic marker Human Epidermal Growth Factor Receptor-2 (HER2) stained tissue sample is presented. The HER2 challenge dataset is used for the study to score the sample under observation. Two CNN networks viz. the Xception network in a transfer learning framework and a proposed simpler CNN architecture AutoIHCNet, with three convolution blocks and dense layers, are used in this study considering 228 x 288 x 3 input shape. The training parameters viz. optimizers, learning rate, and the number of epochs are varied to have 48 sets of experiments to choose the best training settings. From the whole slide image, representative region of interest (ROI) images are extracted. One ROI image is divided into eight sub-image patches. 2400 patches from 300 training ROI images were extracted and out of these 2130 patches are used for training based on stained tissue regions available in the patch. Statistical decision fusion using mode is performed for collective voting from eight sub-image patches to label the sample image under observation. 100 test images are used from different cases, to avoid any bias, to assess the models. The proposed deep learning architectures are also compared with the ImmunoMembrane application. Average test accuracy and Pearson's correlation coefficient are used to assess the performance of automated approaches compared to ground truth. The performance is assessed in terms of improvement in accuracy from the patch-based score to ROI image-based score as well as final comparison for image-based comparison with ImmunoMembrane on 100 separate test images. The architectures, Xception network as transfer learning and AutoIHCNet, with statistical decision fusion, improved the accuracy from 95% to 97% and 96% to 98% respectively for the patch-based score to ROI image-based score whereas, the state-of-the-art ImmunoMembrane application shows 87% accuracy for the ROI image-based score. (C)& nbsp;2022 Elsevier B.V. All rights reserved.
Breast cancerCNNHER2 scoreImmunohistochemistryImage analysisPrognosisCONVOLUTIONAL NEURAL-NETWORKSWHOLE SLIDE IMAGESBREAST-CANCERCELL-MEMBRANESRECOMMENDATIONSCLASSIFICATIONSEGMENTATION
NSTL
Elsevier
